1.1. Installation Instructions¶

This section describes how to set up TI TVM on a host machine for model compilation and on a target EVM for inference.

1.1.1. Prerequisites¶

Host machine (x86_64):

Ubuntu 22.04 LTS
Python 3.10 (virtual environment recommended)
At least 16 GB RAM and 20 GB free disk space

System packages:

sudo apt-get install libyaml-cpp-dev libglib2.0-dev python-setuptools

Target EVM (AArch64):

SD card with Linux file system from appropriate TI Processor Linux SDK installed

1.1.2. Supported Devices¶

Device Family	`SOC` Variable	C7™ NPU Acceleration
J721S2 / TDA4VL	`J721S2` or `TDA4VL`	Yes
J784S4 / TDA4VH	`J784S4` or `TDA4VH`	Yes
J722S / TDA4AEN / AM67A	`J722S` or `TDA4AEN` or `AM67A`	Yes
AM62A	`AM62A`	Yes
AM62	`AM62`	No

1.1.3. Processor SDK Versions¶

TI TVM 0.18.0 is released as part of the following Processor SDK versions:

Device Family	Processor SDK RTOS	Processor SDK Linux
J721E / TDA4VM	11.02.00.06	11.02.00.04
J721S2 / AM68A	11.02.00.06	11.02.00.04
J784S4 / AM69A	11.02.00.06	11.02.00.04
J722S / AM62A / AM62	N/A	N/A

1.1.4. Host Installation (Compilation)¶

The recommended way to install TI TVM is via the edgeai-tidl-tools repository. The setup script installs TVM, TIDL tools, and all required Python dependencies in a single step.

1.1.4.1. Clone edgeai-tidl-tools¶

Clone the repository and check out the tag that matches your SDK version:

git clone https://github.com/TexasInstruments/edgeai-tidl-tools.git
cd edgeai-tidl-tools
git checkout <tag>   # e.g. 11_02_04_00 — see repository tags for your SDK version

1.1.4.2. Run the setup script¶

Note

A Python 3.10 virtual environment is recommended before running the setup script.

./scripts/setup/setup.sh

The setup script installs the following:

Component	Version / Notes
TVM Python wheel	0.18.0 (x86_64, Ubuntu 22.04, Python 3.10)
ONNX Runtime Python wheel	1.15.0 (TI build with TIDL execution provider)
TFLite Runtime Python wheel	2.12.0 (TI build)
TIDL Runtime Python wheel	0.1.0
TIDL tools	Per-SOC libraries downloaded from `software-dl.ti.com`
AArch64 GCC toolchain	arm-gnu-toolchain 13.2.Rel1 (for cross-compilation)
C7x compiler (CGT)	ti-cgt-c7000 5.0.0.LTS (required for `c7x_codegen=1`)
C++ runtime dependencies	ONNX Runtime, TFLite, and TVM C++ headers/libraries
Example models and inputs	Out-of-box data for edgeai-tidl-tools examples

1.1.4.3. Set environment variables¶

Source the environment setup script with your target SoC:

source ./scripts/setup/setup_env.sh <SOC>   # e.g. J721E, J784S4, AM62A

This sets SOC, TIDL_TOOLS_PATH, ARM64_GCC_PATH, and CGT7X_ROOT automatically. See the table below for the full list of variables.

Variable	Description
`SOC`	Target SoC identifier (see Supported Devices above)
`TIDL_TOOLS_PATH`	Path to TIDL libraries (required for TIDL offload)
`ARM64_GCC_PATH`	Path to AArch64 GCC toolchain (required when `compile_for_device=1`)
`CGT7X_ROOT`	Path to C7000 Code Generation Tools (required when `c7x_codegen=1`)

1.1.4.4. Verify the installation¶

python3 -c "import tvm; print(tvm.__version__)"

Expected output: 0.18.0

1.1.5. Target Installation (Inference)¶

All dependencies required to run TVM inference are included in the Linux filesystem packaged with the Processor SDK.

Note

Processor SDK 11.2 only: Two Python packages required by the TVM runtime (psutil and typing_extensions) are missing from this release due to a known issue. Install them manually on the EVM:

pip3 install psutil typing_extensions

This will be resolved in the next SDK release.